Introduction

Dense image-to-3D meshes look great in a viewer and then choke a game engine. ComfyUI-LATO2, shipped by sakalond on GitHub on November 3, 2026 (01:08 UTC), wraps LATO.2 (HUST + Meshy AI, MIT weights) as a ComfyUI custom node that regenerates a clean low-poly mesh — 200 to 5,000 vertices — from dense TRELLIS.2, Hunyuan3D, or scan geometry.

Primary artifact: github.com/sakalond/ComfyUI-LATO2 (GPL-3.0 node; LATO.2 code/weights pulled by the installer). This is an open, runnable game-asset remesh surface inside ComfyUI — not another closed SaaS text-to-3D drop, and not a classical decimator that only deletes triangles.

The Magic: Rebuild Topology, Don’t Just Cull Faces

LATO.2 factorizes mesh generation into a vertex flow (places the vertex budget you ask for on the shape) and a topology flow (connects those vertices). Edges follow the form instead of the input triangle soup. The ComfyUI node then hardens that raw prediction into something you can actually ship downstream:

  • One closed, outward-facing surface — rebuilds faces so edges are not shared by more than two faces, fills holes, and orients normals
  • Inner-wall strip — voxel flood fill drops TRELLIS.2’s sealed inner shell before remeshing
  • Transform preserve — result snaps back to the input’s position and scale for drop-in replacement
  • Native ComfyUI MESH I/O — wires into Preview 3D / Save 3D Model; includes a TRELLIS.2 → LATO.2 example workflow

Hardware bar from the README: NVIDIA GPU ≥ 8 GB VRAM, Windows or Linux, ~13 GB disk for the isolated env + weights (3.4 GB LATO.2 + ~1.1 GB DINOv2). Vendor-claimed runtime: **30–60 s** remesh on RTX 3090 / 4080 (plus 1–3 minutes of CPU if you strip a million-face inner wall).

Pro Tip

Treat LATO.2 as a geometry stage, not a texturing stage. Output is geometry only — no UVs, textures, or materials. Plan a second pass (bake, StableGen, Meshy N1 / Meshy 7 retexture, or your engine UV tools) after you lock the low-poly. For props and characters, start around 1,000–1,500 vertices; bump to ~2,500+ when small hardware detail must survive.

Hands-on: Install Node → Remesh a Dense GLB

  1. Install via ComfyUI Manager (search LATO2), comfy node install comfyui-lato2, or:

cd ComfyUI/custom_nodes
git clone https://github.com/sakalond/ComfyUI-LATO2
cd ComfyUI-LATO2
python install_lato2.py

  1. Restart ComfyUI and run LATO.2 Setup Check (3d/LATO.2). It should report ready.
  2. Open a bundled template under Extensions → ComfyUI-LATO2, or start with example_workflows/lato2_after_trellis2.json if you already have native TRELLIS.2. Set vertex_count (200–5000), keep remove_inner_walls and rebuild_surface on for game-ready solids, and export the GLB from output/lato2/.

Thin parts (blades, slats, cloth edges) can fold — try 2–3 seeds and keep the cleanest hero mesh. Narrow gaps may bridge. Do not expect Boolean-safe or print-safe topology on every seed.

Conclusion

ComfyUI-LATO2 is the practical bridge between “pretty gen mesh” and “engine-tolerable low-poly”: open ComfyUI node, LATO.2 neural remesh at a controllable 200–5,000 vertex budget, TRELLIS.2 / Hunyuan3D-aware cleanup, runnable today on a mid-range NVIDIA card. Research roots: LATO.2 project page / arXiv:2607.10623. For ArtRealmAI’s game-gen beat, the actionable surface is the November 3 node release at GitHub.

Original Source:

https://github.com/sakalond/ComfyUI-LATO2

—Titus